DUSE: A New Benchmark Dataset for Drug User Sentiment Extraction.

Social media continuously produce a huge volume of data in different formats and different domains. In particular, patients' and caregivers' written medical texts play an important role among individuals, medical doctors, and drug developers for understanding drug users' sentiment. However, automatic sentiment detection is a challenging problem in medical settings due to a lack of data with age group, gender, treatment duration, and so on. Therefore, we present a drug review dataset for the most reviewed 100 drugs. Especially, we collected 88K instances from WebMD which is one of the largest online health service providers. Empirically, we explore strongly labeled data and weakly labeled data for automatic sentiment detection using BERT, which learns context-dependent features. We show that the BERT model yields better accuracy than the baseline models.